jnMetaCode/agency-orchestrator

What it is

Agency Orchestrator is a multi-agent AI orchestration engine. You give it a single sentence describing a goal (e.g., "analyze the feasibility of an AI bookkeeping tool"), and it automatically assembles a team of specialized AI "roles" (out of 276 built-in personas like market researcher, financial planner, and UX designer), designs a workflow, runs the steps in parallel where possible, and delivers a complete, structured result.

It's designed for non-programmers: you can use a plain-language CLI command (ao compose "..." --run) or a visual web Studio (ao web), and you can also define workflows manually in simple YAML files.

Core features

• Automatic team formation: The AI picks relevant expert roles from a library of 276 Chinese (and 191 English) personas based on your prompt.
• DAG-based workflow engine: Parses YAML, builds a dependency graph, automatically detects which steps can run in parallel, passes outputs between steps via {{variables}}, and retries failures with exponential backoff.
• Zero-code / low-code: Use a single sentence, or write a simple YAML file with steps, roles, and dependencies.
• Flexible model providers: Works with 20+ APIs. Notably, it can reuse your existing logged-in coding CLIs (Claude Code, GitHub Copilot, OpenAI Codex, etc.) with no API key required, or use traditional keys (DeepSeek, OpenAI, Gemini, etc.). It also supports local models via Ollama.
• Quality gates: Each step can have acceptance criteria (injected into the prompt) and assert rules (mechanical checks like file counts or regex matches) to validate output structure.
• Resume & feedback: Rerun from a checkpoint (--resume) or ask a specific expert to revise its output with your feedback (--feedback).
• Team loadouts: Save a good role lineup as a reusable "team" and apply it to new tasks.
• Prompt Lab: Optimize, test, and version-control your prompts.
• Skills: Attach methodology playbooks (e.g., TDD, structured code review) to specific workflow steps.
• Integrations: Can be called from Cursor, Claude Code, etc. (19 tools), and exposes an MCP server (ao serve).
• Deployment: Installable via npm, or run as a Docker container for NAS/self-hosting.

How it works (simple terms)

1. You type a goal (or write a YAML file).
2. The engine (or the AI composer) picks roles and builds a step-by-step plan (a DAG).
3. Each step calls an LLM with a role-specific system prompt and a task.
4. Steps run in parallel where dependencies allow; outputs are chained.
5. Results are saved to an ao-output/ folder, with options to export to DOCX/PDF or share as a self-contained HTML report.

Who it's for

People who want "a team of AI experts" for planning, research, content creation, code review, or business analysis — without writing Python or managing complex agent frameworks. It's especially friendly for Chinese-language users, given the large built-in Chinese role library.